Validating a model to detect infant crying from naturalistic audio

Human infant crying evolved as a signal to elicit parental care and actively influences caregiving behaviors as well as infant-caregiver interactions. Automated cry detection algorithms have become more popular in recent decades, and while some models exist, they have not been evaluated thoroughly on daylong naturalistic audio recordings. Here, we validate a novel deep learning cry detection model by testing it in assessment scenarios important to developmental researchers. We also evaluate the deep learning model's performance relative to LENA's cry classifier, one of the most commonly used commercial software systems for quantifying child crying. Broadly, we found that both deep learning and LENA model outputs showed convergent validity with human annotations of infant crying. However, the deep learning model had substantially higher accuracy metrics (recall, F1, kappa) and stronger correlations with human annotations at all timescales tested (24 h, 1 h, and 5 min) relative to LENA. On average, LENA underestimated infant crying by 50 min every 24 h relative to human annotations and the deep learning model. Additionally, daily infant crying times detected by both automated models were lower than parent-report estimates in the literature. We provide recommendations and solutions for leveraging automated algorithms to detect infant crying in the home and make our training data and model code open source and publicly available.